The Research Of Voltage Over-Scaling-based Lightweight Hardware Security Technology

It is a challenging task to deploy lightweight security protocols in resource-constrained Io T applications.A hardware-oriented lightweight authentication protocol based on device signature generated during voltage over-scaling(VOS)was recently proposed to address this issue.The computing unit such as adders will produce computing errors in VOS.These computing errors depend on the manufacturing process variation that are difficult to predict,have good uniqueness and reliability,and can be used as the hardware fingerprint of the chip perform identity authentication.However,such as the traditional hardware fingerprint authentication technology—physical unclonable function(PUF),VOLt A is also vulnerable to machine learning(ML)attacks.This paper designs an approximate ML attack model of VOLt A to successfully model and break VOLt A,and proposes a challenge self-obfuscation structure to resist ML modeling attacks.In addition,this paper also designs a VOS-based adversarial examples defense technology,and innovatively uses hardware fingerprint technology to improve the security of neural networks.This paper designs the approximate ML attack model of VOLt A,and successfully breaks VOLt A.VOLt A uses the computing errors generated by the ripple carry adders(RCA)in VOS for identity authentication.Through analyzing the model structure of VOLt A,an approximate attack model of VOLt A is designed,combined with machine learning algorithms completes effective attacks.At the same time,the computing errors generated by RCA in VOS is related to the timing information.The recurrent neural network(RNN)is mainly used to process sequence data,which can better process timing information.This paper uses RNN for modeling for the first time and completes a better attack effect.The experimental results show that artificial neural network(ANN),RNN and covariance matrix adaptation evolution strategy(CMA-ES)can effectively clone the VOLt A challenge-response behavior.Especially,the modeling accuracy of the RNN attack model is as high as 99.65%.This paper proposes a challenge self-obfuscation structure(CSo S).The key idea of CSo S is to combine the previous input with secret keys or random numbers to generate dynamic new keys,and exploit the new keys to obfuscate the current input,than use the challenge correlation enhancement algorithm to enhance the correlation between the previous timing and the current timing challenge to resist the ML attacks.CSo S is not only efficient for VOLt A,but also can be deployed for strong PUFs and exhibits good obfuscation ability.The experimental results show that CSo S-based ML attacks resistant authentication protocol lowers the prediction accuracy of ML on the VOLt A to 51.2%.Furthermore,when we collect 10~6 challenge-response pairs of an Arbiter PUF deployed with CSo S and modeled it using LR,SVM,ANN,RNN and CMA-ES.The experimental results show that modeling accuracy is reduced to 54%.This paper designs a VOS-based technology to enhance the neural network's ability to defend against adversarial examples.The key idea is to use VOS technology to randomize the training set of the neural network model and generate multiple new neural network models,then uses the multiple model to perform integrated classification detection on the confrontation samples to resist the transfer of adversarial examples between different models.The experimental results show that the VOS-based defense technology has defense effects on the adversarial examples generated by FGSM,IGSM,Deep Fool and CW attacks.